Keep Indic virama graphemes together

Author: SqlRush

docs/cc-100-roadmap.md

@@ -1016,6 +1016,8 @@ M7 补充：prompt history `LogEntry` 读取现在接受 `sessionID`/`session`/`
 
 本轮补充：terminal parser 的 text grapheme 分段继续补齐 Unicode mark category 和 Prepend 规则，nonspacing/enclosing/spacing mark 会归入前一 cluster，`का` 这类 spacing mark cluster 不再被拆宽，Arabic prepend mark 加 base text 会保持同一 grapheme，单独 prepend mark flush 时按零宽处理；完整 Unicode UAX #29 分段仍未宣称完成。
 
+本轮补充：terminal parser 的 text grapheme 分段继续补齐常见 Indic virama conjunct，`क्ष` 这类 consonant + virama + consonant cluster 在完整输入和跨 `Feed()` 边界切在 virama 后时都会保持单个窄 grapheme；完整 Unicode UAX #29 分段仍未宣称完成。
+
 本轮补充：terminal CSI parser 现在对 tokenizer flush 出来的非 final-byte incomplete CSI 返回 unknown action，而不是丢弃，贴近官方 `parseCSI` 对 flushed partial sequence 的 fallback 行为。
 
 本轮补充：terminal sequence dispatcher 对 tokenizer flush 出来的 OSC partial sequence 使用 `ParseOSCContent` fallback，允许无 BEL/ST terminator 的 title/link/tab-status content 按官方 parser 语义产出 action。

docs/claude-code-go-rewrite-plan.md

@@ -478,6 +478,7 @@ test/parity/                 # golden tests against TS/official behavior
 - 本轮补充：terminal parser 的 text grapheme 分段继续补齐 Hangul L/V/T jamo 连接规则，decomposed `한` 这类音节在完整输入以及跨 `Feed()` 边界切在 leading/vowel jamo 后时都会保持单个宽 grapheme；完整 Unicode UAX #29 分段仍未宣称完成。
 - 本轮补充：terminal parser 的 text grapheme 分段继续补齐 CRLF line-break cluster，完整输入和跨 `Feed()` 边界切在 `\r`/`\n` 中间时都会保持单个零宽 line-break grapheme，wrap/trim/render 宽度路径也复用同一 line-break 判断；完整 Unicode UAX #29 分段仍未宣称完成。
 - 本轮补充：terminal parser 的 text grapheme 分段继续补齐 Unicode mark category 和 Prepend 规则，nonspacing/enclosing/spacing mark 会归入前一 cluster，`का` 这类 spacing mark cluster 不再被拆宽，Arabic prepend mark 加 base text 会保持同一 grapheme，单独 prepend mark flush 时按零宽处理；完整 Unicode UAX #29 分段仍未宣称完成。
+- 本轮补充：terminal parser 的 text grapheme 分段继续补齐常见 Indic virama conjunct，`क्ष` 这类 consonant + virama + consonant cluster 在完整输入和跨 `Feed()` 边界切在 virama 后时都会保持单个窄 grapheme；完整 Unicode UAX #29 分段仍未宣称完成。
 - 本轮补充：prompt history 写入按官方 `history.ts` 跳过 image pasted content，避免把 image base64/filename/mediaType 存入 `history.jsonl`，读取路径仍兼容旧 image metadata。
 - 本轮补充：paste-cache 增加按 cutoff mtime 清理旧 `.txt` paste 文件的 best-effort 入口，忽略缺失目录、非 `.txt` 文件和单文件错误，和官方 `cleanupOldPastes` 语义对齐。
 - 本轮补充：Buffered prompt history writer 支持撤销最近 pending entry，覆盖官方 `removeLastFromHistory` 在异步 flush 前直接从 pending buffer 移除的 fast path。

docs/first-second-parity-audit.md

@@ -432,6 +432,7 @@ M7 progress now includes:
 - `internal/tui`: terminal grapheme segmentation now applies Hangul L/V/T jamo joining, keeping decomposed syllables such as `한` as one wide grapheme across full input and streaming chunk boundaries.
 - `internal/tui`: terminal grapheme segmentation now keeps CRLF as a single zero-width line-break grapheme across full input and streaming chunks split between `\r` and `\n`, and wrap/trim/render width paths share the same line-break check.
 - `internal/tui`: terminal grapheme segmentation now uses Unicode mark categories for nonspacing/enclosing/spacing marks and handles Prepend characters, keeping clusters such as Devanagari `का` and Arabic prepend-mark plus base text together while treating a lone prepend mark as zero width.
+- `internal/tui`: terminal grapheme segmentation now keeps common Indic virama conjuncts such as Devanagari `क्ष` as one narrow grapheme across full input and streaming chunks split after the virama.
 - `internal/tui`: image hint parsing now accepts OSC ST terminators and base64 `name=` filenames while preserving prompt pasted-content metadata.
 - `internal/session`: prompt history writing now skips image pasted-content records like official `history.ts`, while the reader still accepts older image metadata entries.
 - `internal/session`: paste-cache now has a best-effort cutoff-mtime cleanup helper for old `.txt` paste files, matching official `cleanupOldPastes` behavior.

internal/tui/terminal_parser.go

@@ -274,7 +274,7 @@ func terminalGraphemeMayContinueAtChunkBoundary(value string) bool {
 		return true
 	}
 	last, _ := utf8.DecodeLastRuneInString(value)
-	if last == 0x200d || isTerminalEmojiModifier(last) || terminalGraphemeCanStartKeycapSequence(value) || terminalGraphemeMayContinueHangul(value) {
+	if last == 0x200d || isTerminalVirama(last) || isTerminalEmojiModifier(last) || terminalGraphemeCanStartKeycapSequence(value) || terminalGraphemeMayContinueHangul(value) {
 		return true
 	}
 	regionalCount := 0
@@ -407,6 +407,7 @@ func nextTerminalGrapheme(text string) (string, int) {
 	}
 	prependPrefix := isTerminalPrepend(first)
 	previousWasZWJ := false
+	previousWasVirama := false
 	regionalCount := 0
 	if isTerminalRegionalIndicator(first) {
 		regionalCount = 1
@@ -427,21 +428,30 @@ func nextTerminalGrapheme(text string) (string, int) {
 				}
 			}
 			previousWasZWJ = false
+			previousWasVirama = false
 			continue
 		}
 		if isTerminalCombiningMark(r) || isTerminalSpacingMark(r) || isTerminalVariationSelector(r) || isTerminalEmojiModifier(r) || isTerminalEmojiTag(r) {
 			end += nextSize
 			previousWasZWJ = false
+			previousWasVirama = isTerminalVirama(r)
 			continue
 		}
 		if r == 0x200d {
 			end += nextSize
 			previousWasZWJ = true
+			previousWasVirama = false
 			continue
 		}
 		if previousWasZWJ {
 			end += nextSize
 			previousWasZWJ = false
+			previousWasVirama = false
+			continue
+		}
+		if previousWasVirama && unicode.IsLetter(r) {
+			end += nextSize
+			previousWasVirama = false
 			continue
 		}
 		if regionalCount == 1 && isTerminalRegionalIndicator(r) {
@@ -453,6 +463,7 @@ func nextTerminalGrapheme(text string) (string, int) {
 			end += nextSize
 			hangulClass = terminalHangulJoinedClass(hangulClass, nextHangulClass)
 			previousWasZWJ = false
+			previousWasVirama = false
 			continue
 		}
 		break
@@ -646,6 +657,21 @@ func isTerminalSpacingMark(r rune) bool {
 	return unicode.Is(unicode.Mc, r)
 }
 
+func isTerminalVirama(r rune) bool {
+	switch r {
+	case 0x094d, 0x09cd, 0x0a4d, 0x0acd, 0x0b4d, 0x0bcd, 0x0c4d, 0x0ccd, 0x0d4d, 0x0dca,
+		0x0e3a, 0x0eba, 0x0f84, 0x1039, 0x103a, 0x1714, 0x1734, 0x17d2, 0x1a60, 0x1bab,
+		0x1bf2, 0x1bf3, 0x2d7f, 0xa806, 0xa82c, 0xa8c4, 0xa953, 0xa9c0, 0xaaf6, 0xabed,
+		0x10a3f, 0x11046, 0x11070, 0x1107f, 0x110b9, 0x11133, 0x11134, 0x111c0, 0x11235, 0x112e9,
+		0x112ea, 0x1134d, 0x11442, 0x114c2, 0x115bf, 0x1163f, 0x116b6, 0x1172b, 0x11839, 0x1193d,
+		0x1193e, 0x119e0, 0x11a34, 0x11a47, 0x11a99, 0x11c3f, 0x11d44, 0x11d45, 0x11d97, 0x11f41,
+		0x11f42:
+		return true
+	default:
+		return false
+	}
+}
+
 func isTerminalPrepend(r rune) bool {
 	return (r >= 0x0600 && r <= 0x0605) ||
 		r == 0x06dd ||

internal/tui/terminal_parser_test.go

@@ -239,7 +239,7 @@ func TestTerminalParserDispatchesStringControlActions(t *testing.T) {
 
 func TestTerminalParserSegmentsCommonGraphemeClusters(t *testing.T) {
 	parser := NewTerminalParser()
-	actions := parser.Feed("e\u0301 \u2764\ufe0f 1\ufe0f\u20e3 2\u20e3 \u1112\u1161\u11ab \U0001f44b\U0001f3fd \U0001f469\u200d\U0001f4bb \U0001f1fa\U0001f1f8 \U0001f3f4\U000e0067\U000e0062\U000e0073\U000e0063\U000e0074\U000e007f")
+	actions := parser.Feed("e\u0301 \u2764\ufe0f 1\ufe0f\u20e3 2\u20e3 \u1112\u1161\u11ab \u0915\u094d\u0937 \U0001f44b\U0001f3fd \U0001f469\u200d\U0001f4bb \U0001f1fa\U0001f1f8 \U0001f3f4\U000e0067\U000e0062\U000e0073\U000e0063\U000e0074\U000e007f")
 	if len(actions) != 1 || actions[0].Type != TerminalActionText {
 		t.Fatalf("actions = %#v", actions)
 	}
@@ -254,6 +254,8 @@ func TestTerminalParserSegmentsCommonGraphemeClusters(t *testing.T) {
 		{Value: " ", Width: 1},
 		{Value: "\u1112\u1161\u11ab", Width: 2},
 		{Value: " ", Width: 1},
+		{Value: "\u0915\u094d\u0937", Width: 1},
+		{Value: " ", Width: 1},
 		{Value: "\U0001f44b\U0001f3fd", Width: 2},
 		{Value: " ", Width: 1},
 		{Value: "\U0001f469\u200d\U0001f4bb", Width: 2},
@@ -378,6 +380,21 @@ func TestTerminalParserKeepsChunkedGraphemeClustersTogether(t *testing.T) {
 		t.Fatalf("hangul trailing grapheme = %#v", got)
 	}
 
+	parser = NewTerminalParser()
+	if actions := parser.Feed("\u0915\u094d"); len(actions) != 0 {
+		t.Fatalf("partial virama actions = %#v", actions)
+	}
+	actions = parser.Feed("\u0937!")
+	if len(actions) != 1 || len(actions[0].Graphemes) != 2 {
+		t.Fatalf("virama actions = %#v", actions)
+	}
+	if got := actions[0].Graphemes[0]; got.Value != "\u0915\u094d\u0937" || got.Width != 1 {
+		t.Fatalf("virama grapheme = %#v", got)
+	}
+	if width := TerminalActionsVisibleWidth(actions); width != 2 {
+		t.Fatalf("virama width = %d", width)
+	}
+
 	parser = NewTerminalParser()
 	actions = parser.Feed("\U0001f1fa" + CSISequence(31, "m") + "red")
 	if len(actions) != 2 {